Common chat queries

The fastest way to learn what chat is good at is to see what other founders ask. Every question below is one Herald has answered verbatim or close to it. Copy any of them and adapt to your Product.

Factual questions

The answer is a number, a list, or a name. Herald writes SQL, runs it against your Durable Object, and shows both the answer and the query. Best for “what does the data say?”.

• “How many users signed up last week?”
• “List the ten accounts with the highest usage growth this month.”
• “What’s our MRR today?”
• “Which plan has the lowest churn?”
• “Who’s our oldest paying customer?”

Herald replies with a paragraph + an inline chart or table + the SQL it ran. You can tweak the query in place (“same query but just for accounts > $500 MRR”) and it reruns.

Diagnostic questions

The answer is a reason. Why did something move? What changed between last week and this week? Herald runs three or four exploratory queries in sequence, finds the signal, and narrates it.

• “Why did signups drop on Monday?”
• “Our feature activation rate fell this week — what happened?”
• “Why is our trial-to-paid conversion worse than last quarter?”
• “Which accounts drove our expansion revenue this month?”
• “Why did Bob at Acme churn?”

Diagnostic replies tend to be longer — a paragraph setting up the hypothesis, then one or two queries confirming it, then a single-sentence conclusion. Herald shows its work so you can correct the interpretation.

Exploratory questions

No specific hypothesis. You want Herald to find the shape. These are what chat is best at — they’re too open-ended for a dashboard and too specific for a spreadsheet.

• “What patterns separate our power users from the rest?”
• “Which features correlate with retention after 90 days?”
• “What do accounts that expanded this quarter have in common?”
• “Are there any cohorts silently churning that we haven’t noticed?”
• “What objections came up most in sales calls this month?”

Exploratory answers usually come back as three or four numbered findings with one example each. If one finding is interesting, ask the follow-up: “more on #2” and Herald drills in.

What chat won’t do

Chat doesn’t write content (“draft a launch post”), doesn’t take actions (“send Bob a follow-up”), and doesn’t pull from outside your data (“what’s our market share?”). Every answer is grounded in what you’ve connected — events, revenue, feedback, call transcripts, issues. If Herald can’t answer, it says so and tells you which integration is missing.

Teaching Herald for next time

Every question you ask is logged as an eval example in your Product. If Herald gets an answer wrong, correct it in-line (“that’s the wrong MRR math — treat setup_fee as one-time, not recurring”). The correction becomes a permanent semantic annotation that sharpens the answer forever.

Week 1 of chat is the worst it will ever be. By month three, Herald has your product’s vocabulary.